Blueprint
What is on the Blueprint, and what are the five questions?
Why it works this way
Five questions separate an agent from a script, and the Blueprint gives you an opening answer to each.
Skipping these questions is how an agent ends up with memory, planning and failure handling that nobody chose. The five questions make you look at each one before you spend anything. The answers are openings for you to argue with, which is why they are not an edit box: the point is that the question is asked.
What we gave up: answers you can type over on the page. You cannot edit the five answers.
What it is
The Blueprint appears when you press an opportunity. It is one page, top to bottom:
- The agent's name, a line on what it does, and who it is for. Under that, "Is this worth testing?", where you can answer yes, maybe or no. Your answer changes nothing on your board. It goes to our team. Only the person who started the board sees the question.
- What the agent does.
- The shapes it could ship as, and the button that writes your tests. See What are the ways it could ship, and why pick one?.
- What makes it an agent, not a script, which is the five questions below.
- What might trip it up and What "running well" looks like.
- Why it's buildable now, Who's circling and What's been tried. See Who is already circling, and where does that come from?.
The five questions
A script runs the same steps every time. An agent works out its own. These five decide which one you end up with, and the page gives you an opening answer to each:
- What it remembers. Whether it carries anything from one run to the next, or starts blank every time.
- How it decides the steps.
- What it is allowed to touch. The systems and data it can reach on its own, and the ones it cannot. This is the line between a mistake you can undo and one you cannot.
- How you know the output was good.
- What it does when it fails.
How to use it
- Read the page to the bottom. It is written to be argued with, not accepted.
- Answer each of the five questions for your own agent. If you cannot, you have found something to settle before you build.
- Pick the shape, press Write the tests, then See the tests.
What it will never do
- Tell you the agent will work. "What might trip it up" is a list of ways it could break in use, and it is not a reason to drop the idea.
- Build the agent. Nothing gets built on this page.
- Hide that a scan did not run. If nobody looked, the section says so.
If something looks off
- The page is slow to appear. It is written when you press the opportunity.
- A section says "Nobody has looked yet." The live scan has not run. It runs by itself once the page exists. See the circling guide.
- "The look-up did not come back." The scan failed. There is a button to try it again.
- An older board has no shapes, or no five questions. The page shows what the board has.
- You are a viewer on a shared board. You can read the page, and the "Is this worth testing?" question is not shown to you.